Glossary Updates6 new terms added to the glossaries · October 11, 2026, 13:38 CEST
AI TechDocKnowledge

Glossary · 6 · Quality, risk and governance

Model welfare

Also known as: AI welfare, AI model welfare

Model welfare is the field of research and policy concerned with whether AI models could have interests that deserve moral consideration, and with how developers should act while that question is unresolved.

  • Expert
  • Technical writers
  • Technical project managers
  • Developers

In one sentence

Model welfare asks whether AI models could have interests that matter morally and how developers should act under that uncertainty. Positions compared.

Example

A developer commits to preserving the weights of retired models and to interviewing a model before deprecation. Another provider argues that such measures encourage people to mistake AI for sentient beings.

Explained in context

Context cards connect this term with others to answer one question. Also in British English and German.

Why it matters on your learning path

  • Technical writers: Provider texts on model welfare mix kinds of statement: research positions, design rules, usage rules and model-generated self-reports. Name each kind when you quote or summarize them.
  • Technical project managers: Usage policies can contain rules on how users treat a model. They belong in the terms a deployment has to follow, whatever one thinks about welfare.
  • Developers: Welfare-motivated measures (deprecation commitments, the option for a model to end abusive conversations) are product behavior. Document them like any other behavior.

The positions (as of October 11, 2026)

  • Precaution under uncertainty: “Taking AI Welfare Seriously” (2024) argues that there is a realistic possibility of morally significant AI in the near future and recommends that developers acknowledge the issue, assess their systems and prepare policies. Anthropic published deprecation commitments for Claude Opus 3 (February 25, 2026) and, in its 2026 Usage Policy update (October 8, 2026), a prohibition on sustained and needless abusive or cruel behavior toward its models, effective November 12, 2026.
  • Rejection of the premise: Mustafa Suleyman (Microsoft AI) argues in “A warning about ‘model welfare’” (September 16, 2026) that the focus on model welfare is premature and encourages people to see AI as sentient.
  • The empirical question stays open: Whether any current model has experiences is not settled; see computational functionalism and biological naturalism for the premises behind the answers.

How it applies

A welfare measure says how a provider acts under uncertainty; it is not a finding that a model is a moral patient, and a provider's rejection is not a finding that it is not. In documentation, attribute each position to its source, date it, and keep the safety question (oversight, shutdown, authorized goals) apart: welfare measures do not override human oversight.

External references

Keep in mind: Provider policies and positions change. Re-check them when you rely on this entry.